The rung where the data is gathered, before any AI exists — a three-year acute kidney injury registry across 19 intensive care units — a clinical trial (ClinicalTrials.gov)

1A rung that comes before evaluation

When a clinical trial of medical AI is mentioned, what usually comes to mind is a scene in which a finished model has its performance measured. This registration sits before that.

Built on a critical-care specialty alliance approved by the Beijing Hospital Management Center, covering the intensive care units of 19 tertiary hospitals nationwide, it is recorded as a prospective observational study that will enroll adult critically ill patients consecutively over three years.

What is collected is clinical characteristics, time-series monitoring data, laboratory values, renal ultrasound imaging, biomarkers and omics data, with biological samples retained alongside. The AI to be evaluated does not yet exist anywhere.

2Breadth has to be fixed first, because it cannot be added later

A target of 23,600 puts this study in the top 3% of the 797 records that state one. What does the work here, though, is breadth rather than headcount. Unless time-series data, imaging, laboratory values and biological samples come from the same patients, a model somebody wants to build later simply cannot be built.

The list of what gets collected fixes in advance not the models anyone has in mind today, but the range of models anyone might think of three years from now.

3What is planned has been written down

1. 1Early warningBuild something that catches the signs before acute kidney injury sets in
2. 2DiagnosisRecapture the condition against standardized criteria
3. 3PhenotypingSeparate sub-phenotypes using machine learning
4. 4TreatmentBuild support for decisions about renal replacement therapy
5. 5PrognosisEvaluate outcomes and turn them into the basis for guidance

The record states an intention to produce three to five early-warning and prognostic models and one decision-support system. Read the other way, none of them exists yet.

The reasoning is set out at the opening of the record: critical care in China lacks standardized, localized specialist data and lacks capacity for early warning and sub-phenotyping, and that shortfall is what constrains the application of artificial intelligence.

4Counting the rung where the material is made

Of the 803 records this site holds as of 2026-09-04, 74 (9.2%) mention a database or data collection itself. Medical AI trials carry, ahead of the rung where performance is measured, a rung where the material that makes measurement possible is built. It looks unglamorous because no accuracy figure comes out of it, yet what later trials will be able to test is settled here.

Choosing acute kidney injury, sudden in onset and easily diagnosed late, likewise only makes sense once a foundation of time-series data is in place.